Pith. sign in

REVIEW 6 minor 19 references

On the smallest singular value of the product of random and deterministic matrices

T0 review · 0 major / 6 minor · reviewed 2026-07-10 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Multiplying a random matrix by a fixed invertible matrix replaces the classical 1/√n scale for the smallest singular value by 1 over the Hilbert–Schmidt norm of the inverse.

desk verdict Clean, correct extension of Rudelson–Vershynin invertibility to MA, with the right scale 1/||M^{-1}||_HS and a genuine new projection comparison; fourth-moment bound is the only real soft spot and the authors flag it. read the letter →

arxiv 2607.06785 v1 pith:53WOGM24 submitted 2026-07-07 math.PR

Brayden Letwin , Achintya Raya Polavarapu This is my paper · ORCID
classification math.PR MSC 60B2015B52
keywords smallestsingularvaluerandommatricesHilbert–Schmidtnormcompressibleandincompressiblevectorsanti-concentrationproductofdeterministicstablerank
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

When a random square matrix A with independent mean-zero unit-variance entries is multiplied on the left by a fixed invertible matrix M, the smallest singular value of the product MA is controlled by the Hilbert–Schmidt norm of M inverse, not by the usual 1/√n. The paper proves a lower-tail bound: the probability that this singular value falls below ε times that scale is at most a constant times ε plus an exponentially small term in n. The argument works under a uniform fourth-moment bound and a uniform anti-concentration condition on the entries. In the special case of Gaussian entries the same scale is shown to be sharp for the expectation, so E[s_min(MA)] is comparable to 1/||M^{-1}||_HS. The result therefore gives a natural non-asymptotic replacement for the classical Rudelson–Vershynin invertibility estimate when rows or columns are deterministically rescaled.

What carries the argument

The Loewner comparison E[Π_j] ≼ (C/n) Id for the random orthogonal projection Π_j onto the orthogonal complement of all columns of A except the j-th; this average projection bound supplies the Hilbert–Schmidt factor that appears in the denominator of the distance formula for incompressible vectors.

What would settle it

Produce a sequence of invertible matrices M_n and random matrices A_n whose entries satisfy only second-moment and anti-concentration hypotheses (no uniform fourth-moment bound) for which P(s_min(M_n A_n) ≤ ε / ||M_n^{-1}||_HS) fails to be O(ε) + exp(-c n).

Watch

Extended reading notes

Core claim

Under independent mean-zero unit-variance entries with uniform fourth-moment bound K and uniform Lévy concentration L(a_ij,1) ≤ κ < 1, there exist constants C, c depending only on K and κ such that for every fixed invertible M and every ε ≥ 0 one has P(s_min(MA) ≤ ε / ||M^{-1}||_HS) ≤ Cε + e^{-cn}. For Gaussian A the same scale is sharp in expectation: E[s_min(MA)] ≍ ||M^{-1}||_HS^{-1}.

Load-bearing premise

Every entry of the random matrix is assumed to have a fourth moment bounded by the same constant K; without that uniform bound the key projection comparison fails.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

0 major / 6 minor

Summary. The paper studies the smallest singular value of the product MA, where A is an n imes n random matrix with independent mean-zero, variance-one entries that have uniformly bounded fourth moments and a uniform anti-concentration bound L(a_ij,1)≤κ, and M is a fixed invertible deterministic matrix. The main result (Theorem 1.1) asserts that there exist constants C,c>0 depending only on K and κ such that P(s_min(MA)≤ε/∥M^{-1}∥_HS)≤Cε+e^{-cn} for every ε≥0. The argument follows the classical compressible/incompressible decomposition: the compressible contribution reduces to a cited bound for A itself, while the incompressible contribution is reduced, via a distance formula and a standard large-coordinate lemma, to controlling the random projection Π_j onto the orthogonal complement of all but the j-th column. The novel technical ingredient is Theorem 4.4, which establishes the Loewner comparison c/n Id ⪫ E[Π_j] ⪫ C/n Id by sequential revelation of columns and fourth-moment calculations; this yields an averaged bound on the denominator ∥M^{- op}z_j∥_2. Anti-concentration of the numerator is taken from existing invertibility machinery. In the Gaussian case the authors further prove matching upper bounds (Theorem 5.1, Proposition 5.3) showing that E[s_min(MA)]≍1/∥M^{-1}∥_HS, together with a lower-tail estimate and an example illustrating the necessity of a stable-rank residual term.

Significance. The result supplies the natural scale that replaces the classical 1/√n when a fixed invertible matrix multiplies a random matrix with independent entries. Even the diagonal case already covers row- or column-scaled models whose variances need not be identical, and the Gaussian sharpness confirms that the Hilbert–Schmidt scale is optimal in expectation. The proof is modular and largely self-contained once standard black-box invertibility and anti-concentration statements are granted; the new projection comparison (Theorem 4.4) is of independent interest. The authors candidly flag the fourth-moment hypothesis as the sole obstacle to broader entry distributions, which is a clear and honest limitation rather than a hidden gap. Overall the paper advances the non-asymptotic theory of products of random and deterministic matrices in a clean and usable form.

minor comments (6)
  1. Introduction, paragraph after Theorem 1.1: the remark that results such as [3,10] treat more general entry distributions is helpful; a one-sentence pointer to precisely which moment or anti-concentration assumptions those works use would make the comparison sharper.
  2. Section 2.1: the reduction for finitely many small n via s_min(MA)≥s_min(M)s_min(A) loses a √n factor relative to the HS scale; while harmless for fixed n, a brief remark that the constants may then depend on n (or that one simply absorbs them into C) would avoid any ambiguity.
  3. Theorem 4.4, upper-bound iteration: the choice of d_0 involving (2+2√((K+2)(K+3)))^4 is correct but opaque; a short parenthetical explaining that it forces η_d≤1/2 would improve readability.
  4. Lemma 4.5 and Corollary 4.6: the constants C_{4.5},c_{4.5} are taken from [10]; stating explicitly that they depend only on K and κ (already true) and that the intersection with R_j does not degrade the rate would make the dependence transparent.
  5. Section 5, Proposition 5.5: the construction is clear, but the final comparison 1/∥M^{-1}∥_HS ≤ a_{5.5}/(n√r) could be written with the same a_{5.5} that appears in the definition of M^{-1}, to avoid a momentary notational mismatch.
  6. Typographical: abstract and title use both “V ALUE” and “MA TRICES” with spaces; these should be cleaned for the published version. Occasional double spaces and line-break artifacts (e.g., “W ashington”) appear in the author addresses.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the derivation is a self-contained probabilistic argument with independent black-box inputs and no fitted or self-referential steps.

full rationale

Theorem 1.1 is proved by the standard compressible/incompressible decomposition. Compressible vectors reduce immediately to a known bound on ||Au||_2 (Lemma 3.1 from [9]) via the elementary comparison ||Mx||_2 ≥ ||M^{-1}||_HS^{-1} ||x||_2. Incompressible vectors reduce to controlling dist(MY_j, MH_j) = |⟨z_j, Y_j⟩| / ||M^{- op}z_j||_2; the numerator is handled by anti-concentration from [10], while the novel ingredient is the Loewner comparison E[Π_j] ≺ (C/n) Id (Theorem 4.4), proved by an explicit one-column-at-a-time martingale argument using only fourth-moment Chebyshev bounds and Paley–Zygmund. The Gaussian sharpness section independently recovers the same scale via Edelman tails and Hanson–Wright, without feeding the main theorem back into itself. All external citations ([9], [10], [13], [14], [15], [19]) are used as black-box invertibility or concentration statements whose conclusions do not encode the target scale ||M^{-1}||_HS; none of the authors of the present paper appear among those citations. There are no fitted parameters, no self-definitional normalizations, and no uniqueness theorems imported from the authors. The fourth-moment hypothesis is flagged by the authors as necessary for Theorem 4.4 and is not circular. Score 0 is therefore appropriate.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper works entirely within standard probability and linear algebra. The only domain assumptions are the moment and anti-concentration hypotheses on the entries of A; no free parameters are fitted to data, and no new physical or mathematical entities are postulated. Constants C,c depend only on the input parameters K and κ.

assumptions (4)
  • domain assumption Entries of A are independent with E[a_ij]=0, E[a_ij^2]=1, E[a_ij^4]≤K, and Lévy concentration L(a_ij,1)≤κ for all i,j.
    Standing hypotheses of Theorem 1.1; used throughout Sections 3–4 and for the projection comparison.
  • domain assumption M is a fixed real invertible n×n matrix.
    Hypothesis of the main theorem; the scale is expressed in terms of ||M^{-1}||_HS.
  • standard math Standard probabilistic inequalities (Chebyshev, Cauchy–Schwarz, Paley–Zygmund, Hanson–Wright) and Loewner order for positive-semidefinite matrices.
    Used in the sequential-revelation argument for E[Π_j] (Theorem 4.4) and in the Gaussian section.
  • domain assumption Prior invertibility results for random matrices under the same moment/anti-concentration assumptions ([9, Lemma 5.3], [10, Theorem 1.1 and related lemmas]).
    Invoked for the compressible bound (Proposition 3.2) and for the event that the complement of one column has full rank n−1 (Lemma 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the smallest singular value of the product of random and deterministic matrices." pith.science (2026). https://pith.science/paper/53WOGM24

@misc{pith2026260706785,
  author       = {Pith},
  title        = {Pith review of: On the smallest singular value of the product of random and deterministic matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53WOGM24}},
  note         = {Machine review of arXiv:2607.06785}
}
abstract

Let $A=(a_{ij})$ be an $n\times n$ real-valued random matrix with independent, mean-zero, variance-one entries whose fourth moments are uniformly at most $K$. Suppose that there exists $\kappa \in (0, 1)$ such that the entries of $A$ satisfy $$ \max_{i,j}\sup_{u \in \mathbb{R}} \mathbb{P}(\lvert a_{ij} - u\rvert < 1) \le \kappa. $$ We prove that there are constants $c,C>0$, depending only on $K$ and $\kappa$, such that for every fixed invertible $n\times n$ matrix $M$ and every $\varepsilon\ge0$, $$ \mathbb{P}!\left(s_{\min}(MA) \le \frac{\varepsilon}{\lVert M^{-1}\rVert_{\mathrm{HS}}}\right) \le C\varepsilon + e^{-cn}. $$ In the Gaussian case, we also show that the above estimate is sharp in the sense that $\mathbb{E}[s_{\min}(MA)]\asymp \lVert M^{-1}\rVert_{\mathrm{HS}}^{-1}.$

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

  1. [1]

    Radosł aw Adamczak, Olivier Guédon, Alexander Litvak, Alain Pajor, and Nicole Tomczak-Jaegermann,Smallest singular value of random matrices with independent columns, C. R. Math. Acad. Sci. Paris346(2008), no. 15-16, 853–856. MR 2441920

    work page 2008

  2. [2]

    Math.310(2017), 426–483

    Anirban Basak and Mark Rudelson,Invertibility of sparse non-Hermitian matrices, Adv. Math.310(2017), 426–483. MR 3620692

    work page 2017

  3. [3]

    Max Dabagia and Manuel Fernández,The smallest singular value of inhomogeneous random rectangular matrices, 2024, Preprint, available at arXiv:2408.14389

    work page Pith review arXiv 2024

  4. [4]

    Matrix Anal

    Alan Edelman,Eigenvalues and condition numbers of random matrices, SIAM J. Matrix Anal. Appl.9(1988), no. 4, 543–560. MR 964668

    work page 1988

  5. [5]

    MR 1133335

    ,The distribution and moments of the smallest eigenvalue of a random matrix of Wishart type, Linear Algebra Appl.159(1991), 55–80. MR 1133335

    work page 1991

  6. [6]

    OlivierGuédon, AlexanderE.Litvak, AlainPajor, andNicoleTomczak-Jaegermann,On the interval of fluctuation of the singular values of random matrices, J. Eur. Math. Soc. (JEMS)19(2017), no. 5, 1469–1505. MR 3635358 19

    work page 2017

  7. [7]

    A. E. Litvak, A. Pajor, M. Rudelson, and N. Tomczak-Jaegermann,Smallest singular value of random matrices and geometry of random polytopes, Adv. Math.195(2005), no. 2, 491–523. MR 2146352

    work page 2005

  8. [8]

    Litvak and Omar Rivasplata,Smallest singular value of sparse random matrices, Studia Math.212 (2012), no

    Alexander E. Litvak and Omar Rivasplata,Smallest singular value of sparse random matrices, Studia Math.212 (2012), no. 3, 195–218. MR 3009072

    work page 2012

Show all 19 references
  1. [9]

    Livshyts,The smallest singular value of heavy-tailed not necessarily i.i.d

    Galyna V. Livshyts,The smallest singular value of heavy-tailed not necessarily i.i.d. random matrices via random rounding, J. Anal. Math.145(2021), no. 1, 257–306. MR 4361906

    2021

  2. [10]

    Probab.49(2021), no

    GalynaV.Livshyts, KonstantinTikhomirov, andRomanVershynin,The smallest singular value of inhomogeneous square random matrices, Ann. Probab.49(2021), no. 3, 1286–1309. MR 4255145

    2021

  3. [11]

    heavy-tailed entries, Israel J

    Elizaveta Rebrova and Konstantin Tikhomirov,Coverings of random ellipsoids, and invertibility of matrices with i.i.d. heavy-tailed entries, Israel J. Math.227(2018), no. 2, 507–544. MR 3846333

    2018

  4. [12]

    Mark Rudelson and Roman Vershynin,The least singular value of a random square matrix isO(n−1/2), C. R. Math. Acad. Sci. Paris346(2008), no. 15-16, 893–896. MR 2441928

    2008

  5. [13]

    Math.218(2008), no

    ,The Littlewood-Offord problem and invertibility of random matrices, Adv. Math.218(2008), no. 2, 600–633. MR 2407948

    2008

  6. [14]

    Pure Appl

    ,Smallest singular value of a random rectangular matrix, Comm. Pure Appl. Math.62(2009), no. 12, 1707–1739. MR 2569075

    2009

  7. [15]

    ,Small ball probabilities for linear images of high-dimensional distributions, Int. Math. Res. Not. IMRN (2015), no. 19, 9594–9617. MR 3431603

    2015

  8. [16]

    Szarek,Condition numbers of random matrices, J

    Stanisław J. Szarek,Condition numbers of random matrices, J. Complexity7(1991), no. 2, 131–149. MR 1108773

    1991

  9. [17]

    Vu,Inverse Littlewood-Offord theorems and the condition number of random discrete matrices, Ann

    Terence Tao and Van H. Vu,Inverse Littlewood-Offord theorems and the condition number of random discrete matrices, Ann. of Math. (2)169(2009), no. 2, 595–632. MR 2480613

    2009

  10. [18]

    Theory Related Fields150(2011), no

    Roman Vershynin,Spectral norm of products of random and deterministic matrices, Probab. Theory Related Fields150(2011), no. 3-4, 471–509. MR 2824864

    2011

  11. [19]

    47, Cambridge University Press, Cambridge, 2018, An introduction with applications in data science, With a foreword by Sara van de Geer

    ,High-dimensional probability, Cambridge Series in Statistical and Probabilistic Mathematics, vol. 47, Cambridge University Press, Cambridge, 2018, An introduction with applications in data science, With a foreword by Sara van de Geer. MR 3837109 Department of Mathematics, Uni...

    2018

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.